Research Of Semantic Image Segmentation Based On Fully Convolutional Networks

In the epoch of big data,the quantity of Internet image resources is rising at full speed,and the corresponding processing skill for these images has become a noticeable problem.The major difficulty in image processing is the wide-ranging appliance of image segmentation skill.In the recent few years,the feature extraction capability of deep learning skill makes it have significant research value and widespread application in many aspects such as in computer vision processing,image detection,image semantics segmentation and other kingdoms.Convolutional Neural Networks(CNN)is a pattern of in-depth learning that expresses well in image processing.Fully Convolutional Neural Networks(FCN)which is proposed in recent years is an improved framework of Convolutional Neural Networks.Parallel with conventional Convolutional Neural Networks,Fully Convolutional Neural Networks(FCN)has its characteristic superiority.Structurally,It is similar to the structure of Convolutional Neural Networks.Fully convolution comes true in Fully Convolutional Neural Networks.It means that fully connected layer develops into convolutional layer.It also embodies unique network structures,for instance,the up-sampling layer.Fully Convolution Network can forecast the semantic labels of each and every pixel,realize image semantics segmentation which is at pixel-level,and achieve end-to-end output.Therefore,it can dispose of a great deal of issues of image detection and segmentation.It has obvious strong points compared with traditional neural net.This paper researches the principle of realization of Fully Convolutional Neural Networks(FCN)and Convolutional Neural Networks(CNN),and obtains the dominant position of Fully Convolutional Neural Networks(FCN)through contrast.This paper mainly research the application of Fully Convolutional Neural Networks in the field of image segmentation.Based on the Convolutional Neural Networks which is already existed,it proposes an improved network structure,which tallies with the structure of Fully Convolutional Neural Networks and can achieve end-to-end image segmentation.Output can give consideration to details and situation as a whole through feature fusion.At the same time,this paper designs a low-level network to concentrate on image shape,distance and other features processing,and this network makes the final feature fusion more comprehensive and precise.In this paper,the two main innovation points and homologous advantages are as follows:First,ameliorate the existing Full structure of convolutional networks on the basis of FullyConvolutional Neural Networks,this makes it possible to process images of any size.Meanwhile,it achieves that both input and output are images,and heightens the property of image segmentation.Second,it designs a two-tier network.The first layer is the improved network structure based on fully convolution that is above-mentioned.The second layer is used to concentrate on feature extraction of image details.It mixes the input gray-scale map and depth maps after feature extraction,and the final segmentation result becomes more precise and meticulous.Compared with traditional methods,it has a certain promote.